Improved Algorithms for Computing Fisher's Market Clearing Prices Please share

advertisement
Improved Algorithms for Computing Fisher's Market
Clearing Prices
The MIT Faculty has made this article openly available. Please share
how this access benefits you. Your story matters.
Citation
Orlin, James B. “Improved algorithms for computing fisher’s
market clearing prices.” Proceedings of the 42nd ACM
symposium on Theory of computing, STOC '10, ACM Press,
2010. 291.
As Published
http://dx.doi.org/10.1145/1806689.1806731
Publisher
Association for Computing Machinery
Version
Author's final manuscript
Accessed
Thu May 26 19:33:06 EDT 2016
Citable Link
http://hdl.handle.net/1721.1/68009
Terms of Use
Creative Commons Attribution-Noncommercial-Share Alike 3.0
Detailed Terms
http://creativecommons.org/licenses/by-nc-sa/3.0/
Improved Algorithms for Computing Fisher’s Market
Clearing Prices
∗
James B. Orlin
MIT Sloan School of Management
Cambridge, MA 02139
jorlin@mit.edu
ABSTRACT
We give the first strongly polynomial time algorithm for
computing an equilibrium for the linear utilities case of Fisher’s
market model. We consider a problem with a set B of buyers
and a set G of divisible goods. Each buyer i starts with an
initial integral allocation ei of money. The integral utility
for buyer i of good j is Uij . We first develop a weakly
polynomial time algorithm that runs in O(n4 log Umax +
n3 emax ) time, where n = |B| + |G|. We further modify
the algorithm so that it runs in O(n4 log n) time. These algorithms improve upon the previous best running time of
O(n8 log Umax + n7 log emax ), due to Devanur et al. [5].
Categories and Subject Descriptors
F.2 [Theory of Computation]: Analysis of Algorithms
and Problem Complexity
General Terms
Algorithms, Economics
Keywords
Market Equilibrium, Strongly Polynomial
1.
INTRODUCTION
In 1891, Irving Fisher (see [3]) proposed a simple model of
an economic market in which a set B of buyers with specified
amounts of money were to purchase a collection G of diverse
goods. In this model, each buyer i starts with an initial
allocation ei of money, and has a utility Uij for purchasing all
of good j. The utility received is assumed to be proportional
to the fraction of the good purchased.
In 1954, Arrow and Debreu [2] provided a proof that a
wide range of economic markets (including Fisher’s model)
have unique market clearing prices, that is, prices at which
∗Research supported in part by ONR grant N00014-05-10165.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
STOC’10, June 5–8, 2010, Cambridge, Massachusetts, USA.
Copyright 2010 ACM 978-1-4503-0050-6/10/06 ...$10.00.
the total supply is equal to the total demand. However,
their proof relied on Kakutani’s fixed point theorem, and did
not efficiently construct the market clearing prices. Eisenberg and Gale [8] transformed the problem of computing the
Fisher market equilibrium into a concave cost maximization
problem. They deserve credit for the first polynomial time
algorithm because their concave maximization problem is
solvable in polynomial time using the ellipsoid algorithm.
Devanur et al [5] gave the first combinatorial polynomial
time algorithm for computing the exact market clearing prices
under the assumption that utilities for goods are linear. For
a problem with a total of n buyers and goods, their algorithm runs in O(n8 log Umax +n7 log emax ) time, where Umax
is the largest utility and emax is the largest initial amount
of money of a buyer, and where all data are assumed to be
integral. More precisely, it determines the market clearing
prices as a sequence of O(n5 log Umax + n4 log emax ) maximum flow problems. (For details on solution techniques for
maximum flows, see Ahuja et al. [1]).
Here we provide an algorithm that computes a market
equilibrium in O(n4 log Umax + n3 log emax ) time. We also
show how to modify the algorithm so that the running time
is O(n4 log n), resulting in the first strongly polynomial time
algorithm for computing market clearing prices for the Fisher
linear case.
The running time can be further improved to O((n2 log n)
(m+n log n)), where m is the number of pairs (i, j) for which
Uij > 0. Moreover, the algorithm can be used to provide an
-approximate solution in O((n log 1/)(m + n log n)) time,
which is comparable to the running time of the algorithm
by Garg and Kapoor [10] while using a stricter definition of
-approximation.
1.1
A Formal Description of the Model
Suppose that p is a vector of prices, where pj is the price
of good j. Suppose that x is an allocation of money to
goods; that is, xij is the amount of money that buyer i
spends
P on good j. The surplus cash of buyer i is ci (x) =
ei − j∈G xij . The backorder amount of good j is bj (p, x) =
P
−pj + i∈B xij .
We assume that for each buyer i, there is a good j such
that Uij > 0. Otherwise, we would eliminate the buyer
from the problem. Similarly, for each good j, we assume
that there is a buyer i with Uij > 0. Otherwise, we would
eliminate the good from the problem.
For a given price vector p, the ratio Uij /pj is called the
bang-per-buck of (i, j) with respect to p. For all i ∈ B, we
let αi (p) = max{Uij /pj : j ∈ G}, and refer to this ratio as
the maximum bang-per-buck for buyer i. A pair (i, j) is said
to be an equality edge with respect to p if Uij /pj = αi (p).
We let E(p) denote the set of equality edges.
A solution refers to any pair (p, x) of prices and allocations. We say that the solution is optimal if the following
constraints are all satisfied:
al. [5]) so that prices are always non-decreasing. The algorithm relies on a parameter ∆, called the scaling parameter.
We say that a solution (p, x) is ∆-feasible if it satisfies the
following conditions:
i. ∀i ∈ B, ci (x) ≥ 0;
i. Cash constraints: For each i ∈ B, ci (x) = 0.
ii. ∀j ∈ G, if pj > p0j , then 0 ≤ bj (p, x) ≤ ∆;
ii. Allocation of goods constraints: For each good j ∈ G,
bj (p, x) = 0 .
iii. ∀i ∈ B and ∀j ∈ G, if xij > 0, then (i, j) ∈ E(p), and
xij is a multiple of ∆;
iii. Bang-per-buck constraints: For all i ∈ B and j ∈ G, if
xij > 0, then (i, j) ∈ E(p).
iv. ∀i ∈ B and ∀j ∈ G, xij ≥ 0 and pj ≥ 0.
iv. Non-negativity constraints: xij ≥ 0, pj ≥ 0 for all i ∈
B and j ∈ G.
The above optimality conditions are the market equilibrium conditions specified by Fisher, as well as the optimality
conditions for the Eisenberg-Gale program [8].
1.2
Other Algorithmic Results
The computation of market clearing prices is an important aspect of general equilibrium theory. Nevertheless, as
pointed out in [5], there have been few results concerning the
computation of market equilibrium. We refer the reader to
[5] for these references. Papers that are particular relevant to
this paper include papers by Garg and Karpur [10], Ghiyasvand and Orlin [11], and Vazirani [14]. Garg and Kapoor
present a fully polynomial time approximation scheme (FPTAS) for computing approximate market clearing prices for
the Fisher linear case as well as other cases. Ghiyasvand
and Orlin develop a faster FPTAS while using a stronger
(less relaxed) definition of approximate market equilibrium.
Vazirani develops a weakly polynomial time algorithm for
a generalization of the Fisher linear case in which utilities
are piecewise linear functions of the amounts spent, and in
which there is a utility of saving money. This more general utility function permits the modeling of several types
of constraints, including upper bounds on the amount spent
by buyer i in the purchase of good j.
1.3
Contributions of this Paper
The primary contributions of this paper are a strongly
polynomial time algorithm for computing the market clearing prices, as well as a simple and efficient weakly polynomial time algorithm. Additional contributions of this paper
include the following:
i. An algorithm for obtaining an -approximate solution
that is comparable to the running time of the algorithm
of Garg and Kapoor [10], while simultaneously having
a stricter definition of -approximation.
ii. Our weakly polynomial and strongly polynomial time
algorithms terminate after an optimal set of edges is
identified, which may be much sooner than the termination criterion given by Devanur et al. [5].
1.4
Overview of the Weakly Polynomial Time
Algorithm
The algorithm first determines an initial vector p0 of prices
(as described in Section 3.1). During the algorithm, prices
are modified in a piecewise linear manner (as per Devanur et
We say that a solution (p, x) is ∆-optimal if it is ∆-feasible
and if in addition
v. ∀i ∈ B, ci (x) < ∆.
The ∆-feasibility conditions modify the optimality conditions as follows: buyers may spend less than their initial
allocation of money; it is possible for none of good j to be
allocated if pj = p0j ; more than 100% of a good may be sold;
and allocations are required to be multiples of ∆.
Our algorithm is a “scaling algorithm” as per Edmonds
and Karp [7]. It initially creates a ∆-feasible solution for
∆ = emax . It then runs a sequence of scaling phases. The
input for the ∆-scaling phase is a ∆-feasible solution (p, x).
The ∆-scaling phase transforms (p, x) into a ∆-optimal solution (p0 , x0 ) via a procedure PriceAndAugment, which is
called iteratively. The algorithm then transforms (p0 , x0 )
into a ∆/2-feasible solution. Then ∆ is replaced by ∆/2,
and the algorithm goes to the next scaling phase. One can
show that within O(n log Umax + emax ) scaling phases, when
the scaling parameter is less than 8n−2 (Umax )−n , the algorithm determines an optimal solution.
1.5
Overview of the Strongly Polynomial Time
Algorithm
If xij ≥ 3n∆ at the beginning of the ∆-scaling phase, then
x0ij ≥ 3n(∆/2) for the solution x0 at the beginning of the
∆/2-scaling phase. We refer to edges (i, j) with xij ≥ 3n∆
as “abundant.” Once an edge becomes abundant, it remains
abundant, and it is guaranteed to be positive in the optimal
solution to which the scaling algorithm converges.
If a solution is “∆-fertile” (as described in Section 4) at
the end of the ∆-scaling phase, then there will be a new
abundant edge within the next O(log n) scaling phases. If
the solution (p, x) is not ∆-fertile, then the algorithm calls a
special procedure that transforms (p, x) into a new solution
(p0 , x0 ) that is ∆0 -fertile and ∆0 -feasible for some ∆0 ≤ ∆/n2 .
Since there are at most n abundant edges, the number of
scaling phases is O(n log n) over all iterations.
2.
PRELIMINARIES
We next state how to perturb the data so as to guarantee
a unique optimal allocation. We then establish some results
concerning feasible and optimal solutions.
For a feasible price vector p, it is possible that the set
E(p) of equality edges has cycles. If p is the optimal vector
of prices, this leads to the possibility of multiple optimal
solutions.
Here we perturb the utilities as follows: for all pairs i, j
such that Uij > 0, we replace Uij by Uij + in + j , where
is chosen arbitrarily close to 0. One can use lexicography to simulate perturbations without any increase in worst
case running time. The advantage of perturbations is that
it ensures that the set of equality edges is always cyclefree. Henceforth, we assume that the utilities have been
perturbed, and the algorithm uses lexicography.
If E is the equality graph prior to perturbations, and if
E 0 is the equality graph after perturbations, then E 0 is the
maximum weight forest of E obtained by assigning edge (i, j)
a weight of in + j. (This can be seen if one observes that
(i, j) is an equality edge if log Uij − log pj = log αi ).
In the case that prices are changed and more than one
edge is eligible to become an equality edge (prior to the
perturbation), then the algorithm will select the edge with
maximum weight for the perturbed problem. Henceforth, we
assume that the utilities are perturbed, and we implement
the algorithms using weights as a tie-breaking rule.
Suppose that H ⊆ B × G. If H is cycle free, then the
vector BasicSolution(H) is the unique vector p satisfying
the following:
i. If (i, j) ∈ H and if (i, k) ∈ H, then Uij /pj = Uik /pk .
P
ii. For
i∈C ei =
P each connected component C of H,
j∈C pj .
iii. If (i, j) ∈
/ H, then xij = 0.
P
iv. For all i ∈ B, j∈G xij = ei .
P
v. For all j ∈ G, i∈B xij = pj .
The first two sets of constraints uniquely determine the
vector p. The latter three sets of constraints uniquely determine the allocation vector x. The following lemma implies
that an optimal solution can be obtained from its optimal
set of edges.
∗
Lemma 2. Suppose that (p, x) is any solution and that
B 0 ⊆ B, and G0 ⊆ G. Then
eB 0 − pG0 = cB 0 (x) + bG0 (p, x).
3.
THE SCALING ALGORITHM
3.1
We initialize
as follows. Let ∆0 = emax /n; ∀i ∈ B, let
P
UiG = j∈G Uij ; ∀i ∈ B and ∀j ∈ G, let ρij = Uij ei /nUiG ;
∀j ∈ G, let p0j = max {ρij : i ∈ B}; ∀i ∈ B and ∀j ∈ G,
x0ij = 0.
We note that the initial solution (p0 , x0 ) is ∆0 -feasible.
Moreover, it is easily transformed into a feasible solution by
setting xij = pj for some i with (i, j) ∈ E(p0 ).
3.2
Proof. We first observe that H ∗ is cycle free because the
perturbation ensures that there is a unique optimal solution.
Suppose that (p̂, x̂) = BasicSolution(H ∗ ). The optimal price
vector satisfies the |B| constraints given in the first two sets
of constraints for BasicSolution(H ∗ ). Since these constraints
are linearly independent, p∗ = p̂.
We next consider the remaining three sets of constraints
that determines x̂. For each component C of H ∗ , there
are |C| variables and |C| + 1 constraints, one of which is
redundant because the sum of the moneys equals the sum of
the prices. These constraints are all satisfied by x∗ . Since
there are |H ∗ | variables that are not required to be 0, and
there are |H ∗ | linearly independent constraints, it follows
that x∗ = x̂.
Any solution satisfying the above systems of equations is a
rational number with denominator less than D by Cramer’s
rule. Moreover, if j ∈ G is in component C, then the numerator
of the rational number representing pj is a multiple
P
of i∈B ei , which is at least emin .
In the following, the use of a setP
as a subscript indicates
summation. For example, eS =
i∈S ei . The following
lemma is elementary and is stated without proof.
The Residual Network and Price Changes
Suppose that (p, x) is a ∆-feasible solution at some iteration during the ∆-scaling phase. We define the residual
network N (p, x) as follows: the node set is B ∪ G. For each
equality edge (i, j), there is an arc (i, j) ∈ N (p, x). We refer
to these arcs as the forward arcs of N (p, x). For every (i, j)
with xij > 0, there is an arc (j, i) ∈ N (p, x). We refer to
these arcs as the backward arcs of N (p, x).
Let r be a root node of the residual network N (p, x).
The algorithm will select r ∈ B with cr (x) ≥ ∆. We
let ActiveSet(p, x, r) be the set of nodes k ∈ B ∪ G such
that there is a directed path in N (p, x) from r to k; if k ∈
ActiveSet(p, x, r), we say that k is active with respect to p, x
and r.
As per price changes in [5], we will replace the price pj
of each active good j by q × pj for some q > 1. This is accomplished by the following formula: f (q) = Price(p, x, r, q),
where

fj (q) =
∗
Lemma 1. Suppose that (p , x ) is the optimal solution
for the Fisher Model. Let H ∗ = {(i, j) : x∗ij > 0}. Then
BasicSolution(H ∗ ) = (p∗ , x∗ ). Moreover, if x∗ij > 0, then
x∗ij > emin /D, where D = n(Umax )n .
The Initial Solution
qpj
pj
if j ∈ ActiveSet(p, x, r)
otherwise,
and where q is below a certain threshold.
UpdatePrice(p, x, r) is the vector p0 of prices obtained by
setting p0 = Price(p, x, r, q 0 ), where q 0 is the maximum value
of q such that (p0 , x) is ∆-feasible. At least one of two conditions will be satisfied by p0 : (i) there is an edge (i, j) ∈
E(p0 )\E(p), at which point node j becomes active; or (ii)
there is an active node j with bj (p0 , x) ≤ 0. In the former
case, PriceAndAugment will continue to update prices. In the
latter case, it will carry out an “augmentation” from node
r to node j. The following lemma is straightforward and is
stated without proof.
Lemma 3. Suppose that (p, x) is ∆-feasible and p0 is the
vector of prices obtained by Update-Price(p, x, r). Then (p0 , x)
is ∆-feasible.
3.3
Augmenting Paths and Changes of Allocation
Suppose that (p, x) is a ∆-feasible solution. Any path
P ⊆ N (p, x) from a node in B to a node in G is called
an augmenting path. A ∆-augmentation along the path P
consists of replacing x by a vector x0 where
x0ij
8
< xij + ∆ if (i, j) ∈ P is a forward arc of N (p, x);
xij − ∆ if (i, j) ∈ P is a backward arc of N (p, x);
=
:
xij otherwise.
Lemma 4. Suppose that (p, x) is ∆-feasible and that cr ≥
∆. Suppose further that P is an augmenting path from node
r to a node k ∈ G for which bj (p, x) ≤ 0. If x0 is obtained by
a ∆-augmentation along path P , then (p, x0 ) is ∆-feasible.
Moreover, cr (x0 ) = cr (x) − ∆, and cj (x0 ) = cj (x) for j 6= r.
We will divide the proof of Theorem 1 into several parts.
But first we deal with an important subtlety. We permit a
good j in G to have no allocation if pj = p0j . We need to
establish that node j will receive some allocation eventually,
and that bj (p, x) ≥ 0 at that time.
We present the procedures PriceAndAugment and ScalingAlgorithm next. We will establish their proof of correctness
and running time in Subsection 3.4.
Lemma 5. Suppose that ∆ < p0j /n, and that that (p, x)
is a ∆-feasible solution at the end of the ∆-scaling phase.
Then bj (p, x) ≥ 0.
Algorithm 1. PriceAndAugment(p, x)
begin
select a buyer r ∈ B with cr (x) ≥ ∆;
compute ActiveSet(p, x, r);
until there is an active node j with bj (p, x) ≤ 0 do
replace p by UpdatePrice(p, x, r);
recompute N (p, x), b(p, x), and ActiveSet(p, x, r);
enduntil
let P be a path in N (p, x) from r
to an active node j with bj (p, x) ≤ 0;
replace x by carrying out a ∆-augmentation along P ;
recompute c(x) and b(p, x);
end
Algorithm 2. ScalingAlgorithm(e, U )
begin
∆ := ∆0 ; p := p0 ; x := 0 (as per Section 3.1);
ScalingPhase:
while (p, x) is not ∆-optimal,
replace (p, x) by PriceAndAugment(p, x);
E 0 := {(i, j) : xij ≥ 4n∆};
if BasicSolution(E 0 ) is optimal, then quit;
else continue;
∆ := ∆/2;
for each j ∈ G such that bj (p, x) > ∆,
decrease xij by ∆ for some i ∈ B;
return to ScalingPhase;
end
3.4
Running Time Analysis for each Scaling
Phase and Proof of Correctness
Our bound on the number of calls of PriceAndAugment
during a scaling phase will
P rely on a potential function argument. Let Φ(x, ∆) = i∈B bci (x)/∆c. We will show that
Φ ≤ n at the beginning of a scaling phase, Φ = 0 at the
end of a scaling phase, and each call of PriceAndAugment
decreases Φ by exactly 1.
At the ∆-scaling phase, ScalingAlgorithm selects a buyer r
with cr (x) ≥ ∆. Ultimately it performs a ∆-augmentation.
It continues the ∆-scaling phase until it obtains a ∆-optimal
solution. Each ∆-augmentation reduces cB by ∆. At the
end of the scaling phase, ScalingAlgorithm does a check for
optimality. If an optimal solution is not found, it divides ∆
by 2, obtains a ∆/2-optimal solution and goes to the next
scaling phase.
Theorem 1. During the ∆-scaling phase, ScalingAlgorithm
transforms a ∆-feasible solution into a ∆-optimal solution
with at most n calls of PriceAndAugment. Each call of PriceAndAugment can be implemented to run in O(m + n log n)
time. The number of scaling phases is O(emax +n log Umax );
the algorithm runs in O(n(m + n log n)(emax + n log Umax ))
time.
Proof. We prove the contrapositive of the lemma. Suppose that bj (p, x) < 0. By Property (ii) of the definition
of ∆-feasibility, pj = p0j . Select node i ∈ B such that
(i, j) ∈ E(p0 ). We will show that ci (x) ≥ ∆, which is a
contradiction.
By construction, p0j ≤ ei /n, and so ei /n ≥ n∆. Note that
for each k ∈ G,
Uik
Uij
Uij
Uik
≤ 0 ≤ 0 =
.
pk
pk
pj
pj
Therefore, (i, j) ∈ E(p). Let K = {k ∈ G : xik > 0}. Because (i, k) ∈ E(p), the above inequalities hold with equality. Therefore, for each k ∈ K, pk = p0k .PMoreover, p0k ≤
ei /n for each k ∈ K, and ci (x) = ei − k∈K xik ≥ ei −
P
k∈K (pk + ∆) ≥ ei − |G|ei /n − |G|∆ = |B|ei /n − |G|∆ >
∆.
We now return to the proof of Theorem 1.
Proof of correctness and number of calls of PriceAndAugment. By Lemma 3 and Lemma 4, if a solution is ∆-feasible
at the beginning of a PriceAnd-Augment, then the sequence
of solutions obtained during the procedure are all ∆-feasible.
We next bound the number of calls of PriceAndAugment
during a scaling phase. Lemma 4 implies that each call of
PriceAndAugment reduces Φ by exactly 1.
In the first scaling phase ci (x0 ) ≤ ∆0 for all i ∈ B. Thus,
Φ(x0 , ∆0 ) ≤ |B| < n.
Suppose next that ∆ is a scaling parameter at any other
scaling phase. Let (p, x) be the 2∆-optimal solution at the
end of the 2∆-scaling phase. Thus ci (x) < 2∆ for all i ∈ B.
Therefore, Φ(x, ∆) ≤ |B|. However, immediately after replacing 2∆ by ∆, the algorithm reduces by ∆ the allocation
to any good j with bj (p, x) > ∆. This transformation can
increase Φ by as much as |G|∆. By Lemma 4, the number
of calls of PriceAndAugment is at most |B| + |G| = n.
Proof of time bound per phase. We next show how to
implement PriceAndAugment in O(m+|B| log |B|) time. The
time for identifying an augmenting path, and modifying the
allocation x is O(n). So, we need only consider the total
time for UpdatePrice.
We cannot achieve the time bound if we store prices explicitly since the number of prices changes during PriceAndAugment is O(|B|2 ). So, we store prices implicitly. Suppose
that p is the vector of prices that is the input for PriceAndAugment, and let p0 be the vector of prices at some point
during the execution of the procedure and prior to obtaining
an augmenting path. Let r be the root node of PriceAndAugment, and choose v so that (r, v) ∈ E(p). Rather than
store the vector p0 explicitly, we store the vector p, the price
p0v , and the vector α(p). In addition, for each active node
k, we store γk , which is the price of node v at the iteration
at which node k became active. If node j ∈ G is not active
with respect to p0 , then p0j = pj . If j is active then Then
p0j = pj × p0v /γj .
An edge (i, j) has the possibility of becoming an equality
edge at the next price update only if i is active and j is not
active. Suppose that (i, k) ∈ E(p). In such a case, (i, j)
will become active when the updated price vector p̂ satisfies
Uij /p̂j = αi (p̂) = αi (p)γi /p̂v . Accordingly, if node i ∈ B is
active, and if node j ∈ G is not active, then the price βij of
node v at the iteration at which (i, j) becomes an equality
edge is βij = γi αi (p)pj /Uij .
For all inactive nodes j ∈ G, we let βj = min {βij :
i is active}, and we store the β’s in a Fibonacci heap, a
data structure developed by Fredman and Tarjan [9]. The
Fibonacci heap is a priority queue that supports “FindMin”,
“Insert”, “Delete”, and “Decrease Key”. The FindMin and
Decrease Key operations can each be implemented to run
in O(1) time in an amortized sense. The Delete and Insert
operations each take O(log |B|) time when operating on |B|
elements.
We delete a node from the Fibonacci heap whenever the
node becomes active. We carry out a Decrease Key whenever βj is decreased for some j. This event may occur when
node i becomes active and edge (i, j) is scanned, and βij is
determined. Thus the number of steps needed to compute
when edges become equality edges is O(m + |B| log |B|) during PriceAndAugment.
In addition, one needs to keep track of the price of v at
the iteration at which bj will equal 0. Suppose that node
i ∈ B is active. Let q 0 be the minimum value ≥ 1 such that
bj (q 0 p, x) = 0 mod ∆. Thus bj = 0 when the price of node v
is q 0 γj . Accordingly, βj = q 0 γj , and we store β j in the same
Fibonacci Heap as the inactive nodes of G. Thus the total
time to determine all price updates is O(m+|B| log |B|).
3.5
Abundant Edges and Bounds on the Number of Scaling Phases
Before proving the bound on the number of phases given
in Theorem 1, we introduce another concept and prove a
lemma.
An edge (i, j) is called ∆-abundant if xij ≥ 3n∆, where
x is the allocation at the beginning of the ∆-scaling phase.
(We could use 2n rather than 3n here, but 2n will not be
sufficient for the strongly polynomial time algorithm.)
The concept of abundance was introduced by Orlin [12]
for a strongly polynomial time algorithm for the minimum
cost flow problem. (See also, [13] and Chapter 10 of Ahuja et
al., [1].) The following lemma shows that once an edge (i, j)
becomes abundant, it stays abundant at every successive
iteration. (The reader may note that the following lemma
suggests that it would be better to define “abundance” with
a right hand side of 2∆ rather than 3∆. However, a larger
value is needed for the proof of Property (iv) of Lemma 13.)
Lemma 6. If edge (i, j) is abundant at the ∆-scaling phase,
it is abundant at the ∆/2-scaling phase, and at all subsequent
scaling phases.
Proof. Let x0 and x00 be the initial and final allocations
during the ∆-scaling phase. If (i, j) is ∆-abundant, then
x00ij ≥ 3n∆ − n∆ = 2n∆ = 4n(∆/2). If it is abundant at
the ∆/2-scaling phase, by induction it will be abundant at
all subsequent scaling phases.
Proof of Theorem 1: the number of scaling phases. Suppose that (pk , xk ) and ∆k is the solution and scaling parameter at the beginning of the k-th scaling phase of the
scaling algorithm. Because the change in any allocation and
in any price is at most n∆ during the ∆-scaling phase, the
sequences {pk : k = 1, 2, . . .} and {xk : k = 1, 2, . . .} both
converge in the limit to the optimal solution (p∗ , x∗ ).
Now suppose that ∆k < D/8n, where D = (Umax )−n /n.
Let E k = {(i, j) : xkij > 4n∆k }. We next claim that BasicSolution(E k ) is optimal. Consider edge (i, j). If (i, j) ∈ E k ,
then (i, j) is ∆-abundant, and by Lemma 6, x∗ij > 0. If
(i, j) ∈
/ E k , then x∗ij < xkij + 4n∆k < 1/D. By Lemma
∗
1, xij = 0. We have thus shown that BasicSolution(E k )
is optimal. The number of scaling phases needed to ensure
that ∆k < 1/8nD is O(log emax + n log Umax ).
3.6
Finding Approximate Solutions
Suppose that > 0. A solution (p, x)
P is said to be approximate if it is ∆-optimal for ∆ < j∈G pj .
Theorem 2. . ScalingAlgorithm finds an -approximate
solution after O(log(n/)) scaling phases. The running time
is O(n(m + n log n) log(n/)).
Proof. Recall that ∆0 = emax . Within O(log(n/)) scaling phases, ∆ < ·emax /2n. Suppose (p, x) is the solution at
the end of that phase,
P and let k = argmax{ei : i ∈ B}. Then
ci (x) < ∆ and so j∈G pj > ei − ci (x) > emax /2 > n∆/.
P
Therefore, ∆ < j∈G pj .
The above approach is the fastest method to determine an
-approximate solution. Also, finding an approximate solution is quite practical. For example, if all initial endowments
are in dollars and are bounded by O(f (n)) for some polynomial f (·), then ScalingAlgorithm determines a solution that
is within 1 penny of optimal in O(log n) scaling phases.
3.7
A Simple Strongly Polynomial Time Algorithm if Utilities are Small
Devanur et al. [5] showed how to find an optimal solution
in the case that Uij = 0 or 1 for each (i, j). Here we provide
a simple algorithm for finding an optimal solution if Uij is
polynomially bounded in n.
In the following, we assume that buyers have been reordered so that ei ≥ ei+1 for all i = 1 to |B| − 1. We let
Problem(j) be the market equilibrium problem as restricted
to buyers 1 to j, and including all goods of G. In the following theorem, M = 4n2 (Umax )n . We also relax (iii) of the
definition of ∆-feasibility. We do not require allocations in
abundant edges (see Subsection 3.5) to be multiples of ∆.
Theorem 3. Suppose that ej > M ej+1 , and that k is
the next index after j for which ek > M ek+1 . If one starts
with an optimal solution for Problem(j), ScalingAlgorithm
solves Problem(k) in O(n(k − j) log Umax ) scaling phases.
By solving Problem(i) for each i such that ei > M ei+1 ,
The total number of scaling phases to find an equilibrium
is O(n2 log Umax ).
Proof. Suppose that (p0 , x0 ) is the optimal solution for
Problem(j). The algorithm solves Problem(k) starting with
solution (p0 , x0 ) and with scaling parameter ∆0 = ej+1 . We
first show that this starting point is ∆0 -feasible.
It is possible that allocations will not be a multiple of
∆0 . However, if x0il > 0 for some i ≤ j, then by Lemma 1,
x0il ≥ ej /D > 4nej+1 = 4n∆0 . Thus (i, l) is abundant, and
by Lemma 6, it remains abundant at all subsequent scaling
phases.
We next bound the number of scaling phases for Problem(k). By Lemma 1, the optimum solution (p∗ , x∗ ) for
Problem(k) has the following property. If x∗il > 0 for some
i ≤ k, then x∗il ≥ ek /D. So, the algorithm will stop when
the scaling parameter is less than ek /4nD. Thus, the number of scaling phases is O(log 4nD(ej /ek )) = O(log M k−j ) =
O((k − j)(n) log Umax ).
If we solve Problem(i) for each i such that ei > M ei+1 ,
then the total number of scaling phases to find an equilibrium is O(n2 log Umax ).
4.
A STRONGLY POLYNOMIAL TIME ALGORITHM
In this section, we develop a strongly polynomial time
scaling algorithm, which we call StrongScaling. This algorithm is a variant of the scaling algorithm of Section 3.
The number of scaling phases in StrongScaling is O(n log n),
and the number of calls of PriceAndAugment is O(n) at
each phase. The total running time of the algorithm is
O((n2 log n)(m + n log n)).
4.1
Preliminaries
The algorithm StrongScaling relies on two new concepts:
“surplus” of a subset of nodes and “fertile” price vectors. We
also provide more details about abundant edges.
For the ∆-scaling phase, we let A(∆) denote the set of ∆abundant edges. We refer to each component in the graph
defined by A(∆) as a ∆-component. We let C(∆) denote
the set of ∆-components.
Suppose that p is any price vector during the ∆-scaling
phase. We let N (p, ∆) be the network in which the forward
arcs are E(p) and the backward arcs are the reversals of
∆-abundant edges. We refer to N (p, ∆) as the ∆-residual
network with respect to p. We let ActiveSet(p, ∆, r) be the
set of nodes reachable from r in the network N (p, ∆).
Suppose that H ⊆ B ∪G. We define the surplus
P of H with
respect
to a price vector p to be s(p, H) =
i∈B∩H ei −
P
j∈G∩H pi . Equivalently, s(p, H) = eH − pH .
Suppose that (p, x) is the solution at the beginning of
the ∆-scaling phase. We say that a component H of A(∆)
is ∆-fertile if (i) H consists of a single node i ∈ B and
ei > ∆/3n2 , or if (ii) s(p, H) ≤ −∆/3n2 . We say that the
solution (p, x) is ∆-fertile if any of the components of C(∆)
are ∆-fertile.
Lemma 7. Suppose that (p, x) is the solution obtained at
the end of the ∆-scaling phase. For each fertile ∆-component,
there will be a new abundant edge incident to H within
5 log n + 5 scaling phases.
Proof. Let ∆0 be the scaling factor after at least 5 log n+
4 scaling phases. Then ∆0 < ∆/(16n5 ). Let (p0 , x0 ) be
the solution at the beginning of the ∆0 -scaling phase. If
H = {i}, then ei > 3n2 ∆0 , and so node i is incident to
an edge (i, j) with x0ij > 3n∆0 . If H contains a node of G,
then s(p0 , H) ≤ s(p, H) < −3n2 ∆0 . Because bj (p0 , x0 ) ≥ 0, it
follows that there must be some edge (i, j) with i ∈ B\H and
j ∈ H ∩ G, such that x0ij ≥ 3n∆0 , and so (i, j) is abundant
at the ∆0 -scaling phase.
Because the set of optimal edges is cycle free, there are at
most n − 1 optimal edges, and so there are at most n − 1
abundant edges. If the solution at the beginning of each scaling phase were fertile, then ScalingAlgorithm would identify
all abundant edges in O(n log n) scaling phases, after which
the algorithm would identify the optimal solution. However,
it is possible that a solution is infertile at a scaling phase,
as illustrated in the following example, which shows that
ScalingAlgorithm is not strongly polynomial.
Example. B = {1}; G = {2, 3}; e1 = 1; U12 = 2K ; U13 =
1, and K is a large integer. The unique optimal solution is:
p2 = x12 = 2K /(2K + 1), and p3 = x13 = 1/(2K + 1). At
the k-th scaling phase for k < K, the ∆-optimal solution
has x12 = 1 − 2k and x13 = 2k . It takes more than K scaling phases before (1, 3) is larger than 2n∆ and the optimal
solution is determined.
4.2
The Algorithm StrongScaling
In this section, we present the algorithm StrongScaling. Its
primary difference from ScalingAlgorithm occurs when a solution is ∆-infertile at the beginning of a scaling phase. In this
case, StrongScaling calls procedure MakeFertile, which produces a smaller parameter ∆0 (perhaps exponentially smaller)
and a feasible vector of prices p0 that is ∆-fertile. It then
returns to the ∆0 -scaling phase.
In addition, there are two more subtle differences in the
algorithm StrongScaling. The procedure MakeFertile permits
allocations on abundant edges to be amounts other than a
multiple of ∆. In addition, the procedure produces a solution (p0 , x0 ) such that bj (p0 , x0 ) is permitted to be negative,
so long as bj (p0 , x0 ) ≥ ∆0 /n2 .
Algorithm 3. StrongScaling(e, U )
begin
∆ := ∆0 ; p := p0 ; x := 0; Threshold := ∆;
ScalingPhase:
if BasicSolution(A(∆)) is optimal, then quit;
while (p, x) is not ∆-optimal,
replace (p, x) by PriceAndAugment(p, x);
if no ∆-component is ∆-fertile, and if ∆ ≤ Threshold,
then MakeFertile(p, x)
else do
∆ := ∆/2;
for each j ∈ G such that bj (p, x) > ∆,
decrease xij by ∆ for some i ∈ B;
return to ScalingPhase;
end
Algorithm 4. MakeFertile(p, x, ∆)
begin
∆0 = GetParameter(p, ∆);
if ∆0 > ∆/n2 then Threshold := ∆/n5 and quit;
else continue
p0 := GetPrices(p, ∆0 );
x0 := GetAllocations(p, ∆0 );
end
In the remainder of this section, we let (p, x) be the solution at the end of the ∆-scaling phase, and we assume that
(p, x) is infertile. We use C instead of C(∆).
The procedure MakeFertile(p, ∆) has three subroutines.
The first subroutine is called Get-Parameter(p, ∆). It deter-
mines the next value ∆0 of the scaling parameter. If ∆0 >
∆/n2 , the procedure sets Threshold to ∆/n5 and returns to
the ∆-scaling phase. (The purpose of the parameter Threshold is to prevent MakeFertile from being called for at least
5 log n scaling phases.) In this case, even though p is not ∆fertile, there will be a new abundant edge within O(log n)
scaling phases. If ∆0 ≤ ∆/n2 , MakeFertile continues. The
second subroutine is called GetPrice(p, ∆0 ), which determines a ∆0 -feasible set of prices p0 . The price vector p0 is ∆0 fertile. In addition, −∆0 /2n ≤ s(p0 , H) ≤ ∆0 for all H ∈ C.
The third subroutine is called GetAllocations(p0 , ∆0 ). It obtains an allocation x0 so that (p0 , x0 ) is nearly ∆0 -feasible,
and where allocations are only made on abundant edges.
Subsequently, StrongScaling runs the ∆0 -scaling phase.
In the remainder of this section, we present the subroutines of MakeFertile. In addition, we will state lemmas concerning the properties of their outputs ∆0 , p0 , and x0 . The
algorithms GetParameter and GetPrices each rely on a procedure called SpecialPrice(p, H, t), where H is a ∆-component,
and t ≥ 0. The procedure produces a feasible price vector p̂,
which depends on H and t. The price vector p is increased
until one of the following two events occurs: s(p̂, H) = t, or
there is a component J ∈ C with s(p̂, J) = −s(p̂, H)/2n2 .
The procedure GetParameter computes the next value of
the scaling parameter. For each H ∈ C with at least one
node in G, the procedure runs SpecialPrice(p, H, 0) and computes the vector of prices as well as the scaling parameter,
which we denote as ∆H . If H = {i} (and thus H has no
node of G), then ∆H := ei . The algorithm then sets ∆0 =
max{∆H : H ∈ C}.
The procedure GetPrices computes the next price vector.
It computes SpecialPrice(p, H, ∆0 ) for all H ∈ C, and then
lets p0 be the component-wise maximum of the different price
vectors.
These three procedures are presented next. (We will discuss the procedure GetAllocations later in this section.)
Procedure SpecialPrice(p, H, t)
begin
p0 := p;
select a node v ∈ H ∩ G;
while s(p0 , H) > t and s(p0 , J) > −s(p0 , H)/2n2 ,
for all J ∈ C do
compute the nodes ActiveSet(p0 , ∆, v);
∀ inactive nodes j ∈ G, p̂j := p0j ;
∀ active nodes j ∈ G, p̂j := qp0j ,
where q > 1 is the least value so that
(i) a new edge becomes an equality edge or
(ii) s(p̂, H) = t or
(iii) s(p̂, J) = −s(p̂, H)/2n2 for some J ∈ C;
p0 := p̂;
endwhile
end
Procedure GetParameter(p, ∆)
begin
∀ components H ∈ C do
if H = {i} and i ∈ B, then ∆H := ei ;
else p̂ := SpecialPrice(p, H, 0), and ∆H := s(p̂, H);
∆0 := max {∆H : H ∈ C};
end
Procedure GetPrices(p, ∆0 )
begin
∀ components H ∈ C that contain a node of G do
if s(p, H) ≤ ∆0 , then pH := p;
else pH := SpecialPrice(p, H, ∆0 );
∀j ∈ G, p0j :=max {pH
j : H ∈ C};
end
The following lemma concerns properties of ∆0 and p0
that are needed for the procedure GetAllocation. We let ∆H
and pH denote parameters and prices calculated within GetParameter and GetPrices.
Lemma 8. Suppose that ∆0 and p0 are the output from
GetParameter and GetPrices. Then the following are true:
i. For all H ∈ C, s(p0 , H) ≥ −∆0 /n2 .
ii. If H ∈ C and if H = {i} and i ∈ B, then ei ≤ ∆0 .
iii. For all H ∈ C, s(pH , H) = min{s(p, H), ∆0 }.
0
iv. There is a component H ∈ C , such that pH
H = pH and
0
0
s(p , H) = ∆ .
v. Either there is a node i ∈ B with ei = ∆0 , or there
exists J ∈ C with s(p0 , J) = −∆0 /n2 .
vi. If s(p0 , J) < 0, then there is H ∈ C with s(p0 , H) = ∆0
and J ⊆ ActiveSet(p0 , H).
Proof. This lemma is highly technical, and the proof of
Statements (iv), (v), and (vi) are involved. We first establish
the correctness of Statements (i) to (iii). Prior to Proving
Statements (iv), (v), and (vi), we will establish additional
lemmas.
Proof of parts (i) to (iii). Consider Statement (i). There
exists component K such that s(p0 , H) = s(pK , H) ≥
−s(pK , K)/n2 ≥ −∆0 /n2 , and so the statement is true. Now
consider statement (ii). If H = {i}, then ∆H = ei ≤ ∆0 .
Consider statement (iii). If s(p, H) < ∆0 , then SpecialPrice(p, H, ∆0 ) lets pH = p. Otherwise, SpecialPrice increases the prices, and thus decreases the value s(pH , H),
until s(pH , H) = ∆0 . The other termination criteria do not
play a role until s(pH , H) = ∆H ≤ ∆0 .
4.3
The Auxiliary Network and Maximum Utility Paths
Before proving Statements (iv) to (vi) of Lemma 8, we
develop some additional properties of prices and paths. We
again assume that (p, x) is the input for MakeFertile. Let
A(∆) be the set of abundant edges. The auxiliary network
is the directed graph N ∗ with node set B ∪ G, and whose
edge set is as follows. For every i ∈ B and j ∈ G, there is
an arc (i, j) with utility Uij . For every (i, j) ∈ A(∆), there
is an arc (j, i) with utility Uji = 1/Uij . (There will also be
∗
an arc (i, j).) For every directed
Q path P ∈ N , we let the
utility of path P be U (P ) = (i,j)∈P Uij . For every pair of
nodes j and k, we let µ(j, k) = max{U (P ) : P is a directed
path from j to k in N ∗ }. That is, µ(j, k) is the maximum
utility of a path from j to k.
Lemma 9. The utility of each cycle in the auxiliary network N ∗ is at most 1.
Proof. Let C be any cycle in the auxiliary network.
0
For each edge (i, j) of the auxiliary network, let Uij
=
0
Uij /(αi (p)·pj ). Since Uij /pj ≤ αi (p), it follows that Uij ≤ 1
for all (i, j). Q
0
Let U 0 (P ) = (i,j)∈P Uij
. It follows that U (P ) = U 0 (P ) ≤
1, and the lemma is true.
We assume next that all components of A(∆) have a single
root node, and that the root node is in G whenever the
component has a node of G. We extend µ to being defined
on components as follows: µ(H, K) = µ(v, w), where v is
the root node of H, and w is the root node of K.
Lemma 10. Suppose that K ∈ C and that pK = SpecialPrice(p, K, t). If component H ⊆ ActiveSet(pK , ∆, K), then
K
µ(K, H) = pK
H /pK .
0
Proof. For each (i, j) in the auxiliary network, let Uij
=
K
K
0
Uij /(αi (p ) · pj ). Then Uij ≤ 1 for all (i, j). Suppose
next that P is any maximum utility path from a node v to
a node w with respect to U . Then P is also a maximum
utility path from v to w with respect to U 0 . Moreover, if
K
v ∈ G and w ∈ G, then U 0 (P ) = U (P ) · pK
v /pw .
We now consider the path P 0 from the root of K to the
root of H in the residual network N (pk , ∆). Such a path
exists because H is active. Then each arc of P 0 is an equality
arc, and U 0 (P 0 ) = 1. It follows that P 0 is a maximum utility
K
K
K
path. Moreover, U (P ) = U 0 (P ) · pK
H /pK = pH /pK .
Lemma 11. Suppose that K and H are components of C,
and that pK and pH are price vectors determined in GetK
H
K
Prices. If pH
H < pH , then pH /pK < µ(H, K).
K
K
K
H
K
By assumption,
Proof. pH
H /pK = pH /pH × pH /pK .
K
K
,
<
1.
This
implies
that
p
<
p
/p
pH
H
H and so compoH
H
K
nent H ⊆ ActiveSet(pK , ∆, K). By Lemma 10, pK
H /pK =
K
<
µ(H,
K).
/p
µ(H, K). Therefore pH
K
H
4.4
The Rest of the Proof of Lemma 8
We are now ready to prove Statement (iv), (v), and (vi)
of Lemma 8.
Proof of Statement (iv) of Lemma 8. We assume that
there is no component H such that s(p0 , H) = ∆0 , and we
will derive a contradiction. By statement (iii), if s(p, H) ≥
∆0 , then s(pH , H) = ∆0 . By construction, if s(pH , H) = ∆0
and if s(p0 , H) < ∆0 , then there is some component K such
K
that s(pK , K) = ∆0 , and pH
H < pH . In such a case, we say
that K dominates H.
Let H1 be a component for which s(pH , H) = ∆0 . By our
assumption, there is a sequence of components H1 , H2 , . . . , Hk
such that Hi+1 dominates Hi for i = 1 to k−1, and Hk dominates H1 . For each component J ∈ C, let g(J) = pJJ . Let
Hk+1 = H1 . By Lemma 11,
k
Y
i=1
µ(Hi+1 , Hi ) >
k
Y
g(Hi )/g(Hi+1 ) = 1.
i=1
For i = 1 to k, let Pi be the max multiplier path from
Hi+1 to Hi , and let P = P1 , P2 , . . . , Pk be the closed walk
obtained by concatenating the k paths. Then U (P ) =
Qk
i=1 µ(Hi+1 , Hi ). But by Lemma 9, U (P ) ≤ 1, which is
a contradiction. We conclude that there is a component H
such that s(p0 , H) = ∆0 .
Proof of Statement (v) of Lemma 8. Choose H such that
s(p0 , H) = ∆0 . If H = {i}, ∆0 = ei . The remaining case in
when H contains a node of G. Then the termination criteria for SpecialPrice(p, H, 0) specifies that that there is a
component J such that s(pH , J) = −∆0 /n2 . We claim that
s(p0 , J) = s(pH , J) = −∆0 /n2 . Otherwise, there is a compoH
K
H
nent K such that pK
J > pJ . But then s(p , J) < s(p , J) =
H
2
0
2
K
2
−s(p , H)/n = −∆ /n ≤ −s(p , K)/n , contradicting
the termination rule for SpecialPrice(p, K, ∆0 ).
Proof of Statement (vi) of Lemma 8. We first prove the
statement in the case that s(p0 , J) < s(p, J), and thus p0J >
pJ . In this case, there exists a component H such that
p0J = pH
J > pJ . Let Q be the max utility path from H to
J in N ∗ . We claim that p0H = pH
H . Otherwise, there is a
H
component K that dominates H; that is, pK
H > pH . By
H
K
H
Lemma 10, pK
>
p
.
Therefore,
s(p
,
J)
<
s(p
, J) =
J
J
−s(pH , H)/n2 = −∆0 /n2 ≤ −s(pK , K)/n2 , a contradiction.
Thus Statement (vi) is true in this case.
We will show that there is a component H with s(p, H) ≥
n∆0 and J ⊆ Activeset(p, ∆, H). In this case, p0J ≥ pH
J > pJ ,
and thus Statement (vi) is true because it reduces to the case
given above. We next determine an allocation x̂ as restricted
to the abundant edges and that is nearly ∆-feasible. We then
use flow decomposition (see Ahuja et al. [1] pages 79-81) to
determine the component H.
Consider the allocation x̂ obtained from p that satisfies
the following:
i. if (i, j) ∈
/ A(∆), then x̂ij = 0;
ii. c(x̂) ≥ 0; if s(p, H) ≥ 0, then cH (x̂) = s(p, H);
iii. b(p, x̂) ≤ 0; if s(p, H) < 0, then bH (p, x̂) = s(p, H).
Consider the flow decomposition of y = x − x̂. The flow
decomposition is the sum of nonnegative flows in paths in
N (p, ∆),Pwhere each path is from a deficit node (a node
vPwith
k∈G ŷvk > 0) to an excess node (a node w with
k∈B ŷkw < 0).
Consider the flow in the decomposition out of component
J. We claim that J is a deficit component, and the deficit is
at least (n − 1)∆/n. To see that this claim is true, note that
bJ (p, x) ≥ 0, and bJ (p, x̂) = s(p, J) < 0. Moreover, since
xij is a multiple of ∆ on non-abundant edges, it follows that
bJ (p, x) ≡ s(p, J) mod ∆, and thus bJ (p, x) ≥ s(p, J) +
∆ > −∆/n2 + ∆ > (n − 1)∆/n. Since bJ (p, x̂) ≤ 0, it
follows that there must be at least (n − 1)∆/n units of flow
into component J in the flow decomposition. Accordingly,
there must be at least one excess component H such that
the flow from H to J in the flow decomposition is at least
∆/n. Since all flows in the decomposition are on equality
edges, it follows that J ⊆ ActiveSet(p, ∆, H). Moreover,
s(p, H) ≥ ∆/n ≥ n∆0 , which is what we needed to prove.
Therefore (vi) is true.
4.5
Determining the Allocations
In procedure GetAllocations, for each ∆-component H with
at least two nodes, we designate one node in H ∩ B as the
B-root of H, and we designate one node in H ∩ G as the
G-root of H. If H has a single node k, then k is the root.
The procedure will determine the allocation x0 by first
obtaining a basic feasible solution for a minimum cost flow
problem. (See, e.g., [1].) In the procedure, we will let dk
denote the supply/demand of node k. If i ∈ B, then di ≥ 0.
If j ∈ G, then dj ≤ 0.
Procedure GetAllocations(p0 , ∆0 )
begin
∀ components H ∈ C
if i is the B-root of H, then di = max {0, eH − p0H };
if j is the G-root of H, then dj = min {0, eH − p0H };
if k is not a B-root or G-root, then dk = 0;
select x0 as the unique solution satisfying
the supply/demand constraints and such that
x0ij = 0 for all non-abundant edges (i, j);
end
4.6
On the Correctness of StrongScaling and
its Running Time
In this section, we provide lemmas that lead up to the theorem stating the correctness and time bounds for StrongScaling.
Lemma 12. Suppose that ∆0 is the parameter obtained by
running GetParameter, and suppose that ∆0 > ∆/n2 . If the
Scaling-Algorithm is run starting with solution (p, x) and with
the ∆-scaling phase, then there will be a new abundant edge
within O(log n) scaling phases.
Proof. By Lemma 8, we can select a component H such
that s(pH , H) = ∆0 . Then there must also be a component
J with s(pH , J) = −∆H /n2 . (The termination criterion
would not end with ei = ∆0 because (p, x) was not fertile.)
Because ∆0 > ∆/n2 , StrongScaling went to the ∆/2-scaling
ˆ <
phase, starting with the solution (p, x). Suppose that ∆
∆0 /n2 , and let (p̂, x̂) be a feasible solution at the end of
ˆ
the ∆-scaling
phase. We will show that (p̂, x̂) or else (p̂, x̂)
ˆ
is ∆-fertile, and there will be a new abundant edge within
O(log n) scaling phases. We consider two cases. Assume
0
2ˆ
first that p̂H ≤ pH
H . In this case, s(p̂, H) ≥ ∆ ≥ n ∆, and
there must be an abundant edge directed out of H.
In the second case, p̂H > pH
H . In this case, s(p̂, J) ≤
ˆ 4 . In this case, J is ∆-fertile,
ˆ
s(pH , J) = −∆0 /n2 < −∆/n
and there will be a new abundant edge in O(log n) scaling
phases.
Lemma 13. Suppose that ∆0 and (p0 , x0 ) are the parameter and solution obtained by running MakeFertile, and suppose that ∆0 ≤ ∆/n2 . Then the following are true:
i. (p0 , x0 ) is ∆0 -fertile.
ii. ci (x0 ) ≤ ∆0 for all i ∈ B.
iii. −∆0 /n2 ≤ bj (p0 , x0 ) ≤ 0 for all j ∈ G.
iv. If (i, j) ∈ A(∆), then x0ij > 2n∆0 .
v. If ∆00 < ∆/3n2 , then the solution at the ∆00 -scaling
phase is ∆00 -feasible. Within an additional 2 log n + 4
scaling phases, there will be a new abundant edge.
Proof. Statement (i) follows from Statements (iv) and
(v) of Lemma 8. Statements (ii) and (iii) follow from Statements (i) and (iii) of Lemma 8, as well as the way that x0 is
constructed.
We next prove Statement (iv). Consider x̂ described in
the proof of Statement (vi) of Lemma 8. We first claim
that cB (x̂) < (n + 1)∆. The claim is true because cB (x̂) +
bG (p, x̂) = eB −pG = cB (x)+bG (p, x) ≤ n∆. Also, bj (p, x̂) ≥
−∆/n2 . Thus cB (x̂) ≤ n∆ − bG (p, x̂) < n∆ + (|G|/n2 )∆ <
(n + 1)∆.
Now consider the flow decomposition of y = x − x̂ in the
proof of Statement (vi) of Lemma 8. The sum of the excesses
is at most cB (x̂) < (n + 1)∆. It follows that the total flow
in the decomposition emanating from excess nodes is less
than (n + 1)∆. Therefore, |x̂ij − xij | < (n + 1)∆ for all
(i, j) ∈ A(∆).
Now consider x0Pas determined
P by Procedure
P GetAlloca0
tions. Note that
j∈G pj ≤
j∈G pj ≤
j∈G pj + n∆.
Therefore, |x̂ij − x0ij | < n∆ for all (i, j) ∈ A(∆). Thus
|xij − x0ij | < (2n + 1)∆. Because xij ≥ 3n∆, x0ij > 3n∆ −
(2n + 1)∆ = (n − 1)∆ > (n3 − n)∆0 , and thus (i, j) is abundant at the ∆0 -scaling phase.
We next prove Statement (v). Let (p00 , x00 ) be the solution
at the ∆00 -scaling phase, where ∆00 < ∆0 /4n2 . Let J be
a component with s(p0 , J) < 0. Statement (v) says that
bJ (p00 , x00 ) ≥ 0.
So, suppose instead that bJ (p00 , x00 ) < 0. We will derive a
contradiction. We note that p00J = p0J . (Recall that a node j
only increases its price when j is active and when there is no
augmenting path. This cannot occur for a node j with bj <
0.) We next claim that p00k = p0k if J ⊂ ActiveSet(p0 , ∆, H).
To see why, let Q0 be a maximum multiplier path from k
to J. Then by Lemma 10, p0J = p0k · µ(k, J), and p00J ≥
p00k · µ(k, J). Since p0J = p00J , and p00k > p0k , it follows that
p00k = p0k .
By Statement (vi) of Lemma 8, there is a component H
with s(p0 , H) = ∆0 . In addition, J ⊆ ActiveSet(p0 , ∆, H).
We are now ready to show to derive the contradiction, and
thus show that bJ (p00 , x00 ) ≥ 0. Note that s(p00 , H) = ∆0 ,
and b(p00 , H) ≤ ∆00 ≤ ∆/4n2 . Now consider the flow decomposition y 0 of x00 − x0 . The flow out of component H in y 0
is at least (n − 1)∆0 /n, and so there must be a flow of at
least ∆0 /n to some component K. Each arc in the flow decomposition is in N (p0 , x0 ). Then p00K = p00H · µ(H, K), which
implies that p00K = p0K . But then bK (p00 , x00 ) = bK (p0 , x00 ) >
bK (p0 , x0 )+∆0 /n > (4n−1)∆00 , which is a contradiction.
Theorem 4. Optimality of Strong Scaling. The algorithm StrongScaling finds an optimal solution (p∗ , x∗ ) in
O(n log n) scaling phases. It calls MakeFertile at most n
times, and each call takes O(n(m + n log n)) time. It calls
PriceAndAugment O(n) times per scaling phase. It runs in
O((n2 log n)(m + n log n) time.
Proof. Once all of the abundant edges have been identified, the algorithm finds the optimal solution. And a new
abundant edge is discovered at least once every O(log n) scaling phases. Thus the algorithm terminates with an optimal
solution.
We now consider the running time. The algorithm calls
MakeFertile at most n times because each call leads to a
new abundant edge. Each call of MakeFertile takes O(n(m +
n log n) time. There are O(log n) scaling phases between
each call of MakeFertile, and each scaling phase takes O(n(m+
n log n) time. Thus the running time of the algorithm is
O((n2 log n)(m + n log n)) time.
5.
SUMMARY
We presented a new scaling algorithm for computing the
market clearing prices for Fisher’s linear market model. We
anticipate that this technique can be modified to solve extensions of the linear model.
We have also presented a strongly polynomial time algorithm. The strongly polynomial time algorithmic development may possibly be modified to solve extensions of the
linear model; however, such modifications are challenging
because of the technical aspects of the proof of validity.
6.
ACKNOWLEDGEMENTS
The author gratefully acknowledges all of the help and advice from Dr. Mehdi Ghiyasvand.
7.
REFERENCES
[1] R.K. Ahuja, T.L. Maganti, and J.B. Orlin. Network
Flows: Theory, Algorithms, and Applications. Prentice
Hall, N.J. 1993.
[2] K. Arrow and G. Debreu. Existance of an equilibrium
for a competitive economy. Econometrica 22:265–290,
1954.
[3] W.C. Brainard, H.E. Scarf, D. Brown, F. Kubler, and
K.R. Sreenivasan. How to compute equilibrium prices
in 1891. Commentary. The American Journal of
Economics and Sociology 64: 57-92, 2005.
[4] D. Chakrabarty, N. Devanur, and V. Vazirani. New
results on rationality and strongly polynomial
solvability in Eisenberg-Gale markets. In Proceedings
of 2nd Workshop on Internet and Network Economics,
239-250, 2006.
[5] N.R. Devanur, C.H. Papadimitriou, A. Saberi, and
V.V. Vazirani. Market equilibrium via a primal–dual
algorithm for a convex program. Journal of the ACM
55: 1-18, 2008.
[6] N.R. Devanur and V.V. Vazirani. The spending
constraint model for market equilibrium: algorithmic,
existence and uniqueness results. Proceedings of 36th
Symposium on the Theory of Computing, 519-528,
2004.
[7] J. Edmonds., and R.M. Karp. Theoretical
improvements in algorithmic efficiency for network
flow problems. Journal of ACM 19: 248-264, 1972.
[8] E. Eisenberg and D. Gale. Consensus of subjective
probabilities: the Pari-Mutuel method. Annals of
Mathematical Statistics 30: 165-168, 1959.
[9] M.L. Fredman, and R.E. Tarjan. Fibonacci heaps and
their uses in improved network optimization
algorithms. Journal of ACM 34: 596-615, 1987.
[10] R. Garg and S. Karpur. Price Roll-Backs and Path
Auctions: An Approximation Scheme for Computing
the Market Equilibrium Lecture Notes in Computer
Science, Volume 4286/2006 Springer Berlin /
Heidelberg 225-238, 2006.
[11] M. Ghiyasvand and J.B. Orlin An improved
approximation algorithm for computing Arrow-Debreu
market prices MIT Technical Paper. Submitted for
publication. 2009.
[12] J.B. Orlin. Genuinely polynomial simplex and
non-simplex algorithms for the minimum cost flow
problem. Technical Report 1615-84, Sloan School of
Management, MIT, Cambridge, MA. 1984.
[13] J.B. Orlin. A faster strongly polynomial minimum cost
flow algorithm. Operations Research 41: 338-350, 1993.
[14] V.V. Vazirani. Spending constraint utilities, with
applications to the Adwords market. Submitted to
Math of OR, 2006.
[15] V.V. Vazirani. Combinatorial Algorithms for Market
Equilibrium. in N. Nisan, T. Roughgarden, E’. Tardos,
and V. Vazirani, editors, Algorithmic Game Theory,
pages 103-134, Cambridge University Press,
Cambridge, UK., 2007.
Download